ARTIFICIAL NEURAL NETWORKS LEARNING IN ROC SPACE

In order to control the trade-off between sensitivity and specificity of MLP binary classifiers, we extended the Backpropagation algorithm, in batch mode, to incorporate different misclassification costs via separation of the global mean squared error between positive and negative classes. By achieving different solutions in ROC space, our algorithm improved the MLP classifier performance on imbalanced training sets. In our experiments, standard MLP and SVM algorithms were compared to our solution using real world imbalanced applications. The results demonstrated the efficiency of our approach to increase the number of correct positive classifications and improve the balance between sensitivity and specificity.